Deep Representation Learning on Whole-Brain Population Dynamics Uncovers Geometrically Separable Neural Codes

Extracting interpretable representations from high-dimensional whole-brain neural dynamics remains a major computational challenge in systems neuroscience. Here, we develop a wiring-agnostic deep-learning framework that combines convolutional encoding with temporal transformers to learn compact representations directly from volumetric calcium imaging of the entire \textit{Drosophila melanogaster} brain, without neuronal identification or anatomical annotation. Applied to brain-wide activity recorded across 16 factorially combined sensory and internal-state conditions, the learned representations revealed a factorized organization of whole-brain dynamics: metabolic state, sensory modality, and stimulus valence emerged along three near-orthogonal axes from a classification objective based only on flat class labels and without explicit disentanglement constraints. Spatial attribution and brain region-level ablation analyses linked modality representations to anatomically distinct circuits, whereas state- and valence-related information showed a broadly distributed organization. Our approach provides a scalable framework for discovering interpretable representations of whole-brain neural dynamics.